Prompt Injection Defenses
Work
WorkIn progress
Standalone defense package with 4 layers: constitutional reasoning (principal hierarchy), input sanitization (20+ regex injection patterns, unicode sanitizer, homograph detector), tool governance (per-tool CLI allow/deny lists, tiered file edit guard), and output scanning (secret detection, PII scanning). Features multi-turn trust decay detection, deterministic CLI guard evaluation with wrapper normalization, and configurable allowlists. MIT licensed.